MG205: Econometrics Theory and Applications

Empirical Exercise 5: Beauty

Jose Ignacio Gonzalez-Rojas
j.i.gonzalez-rojas@lse.ac.uk

London School of Economics and Political Science

January 26, 2026

Today: Decomposing the Beauty Premium

From Correlation to Mechanism

Roadmap

  1. Observational evidence: The 15% beauty premium
  2. Four competing channels
  3. Experimental design: Mobius & Rosenblat (2006)

Goals

  1. Test each channel separately
  2. Quantify contributions

Beauty Pays

Attractive Workers Earn 10-15% More Than Average

Hamermesh & Biddle (1994)

Stylized Facts

  • Premium persists across occupations and industries
  • Effect remains after controlling for education, experience, occupation
  • Magnitude similar to gender wage gap

The Puzzle

If beauty doesn’t affect productivity, why would employers pay more?

Could you profit by hiring less attractive workers at lower wages?

Below-Average Looks Cost 15% in Wages

Equivalent to Losing 2.5 Years of Education

Effect of below-average looks vs. one year of education on log wages. Error bars show 95% confidence intervals.

Being rated below-average costs as much as missing 2.5 years of education

Four Channels Could Explain the Beauty Premium

Only One Is Economically Efficient

Gray = Productive (efficient) | Gold = Potentially discriminatory

Channels Have Policy Implications

The Wage Equation Separates Each Channel

Interactions Identify Visual and Oral Effects

Wage Equation

\[\begin{aligned} \log(\text{wage})_{ij} &= \beta_0 + \beta_1 \text{beauty}_j + \beta_2 (\text{beauty}_{j} \cdot \underbrace{\mathbb{1}[i \text{ sees } j\text{'s photo}]}_{\text{Visual}}) \\ &+ \beta_3 (\text{beauty}_j \cdot \underbrace{\mathbb{1}[i \text{ hears } j\text{'s voice}]}_{\text{Oral}}) + \gamma X_{ij} + \varepsilon_{ij} \end{aligned}\]

Confidence Equation

\[\begin{aligned} \log(\text{confidence})_j &= \alpha_0 + \alpha_1 \text{beauty}_j + \alpha_2 \log(\text{actual performance})_j \\ &+ \alpha_3 \log(\text{projected performance})_j + \delta X_j + \nu_j \end{aligned}\]

If \(\alpha_1 > 0\) after controls \(\Rightarrow\) overconfidence

Different Channels Require Different Policy Responses

Efficiency vs. Discrimination

If Productivity Channel

  • Premium is efficient
  • Beautiful workers are actually more productive
  • No intervention needed
  • Hiring based on looks is rational

If Stereotype/Confidence Channels

  • Premium reflects bias
  • Beautiful workers are not more productive
  • Blind hiring could help
  • Employer training may be needed

We need an experiment to separate these channels

An Experiment Can Separate Channels

Mobius & Rosenblat (2006) Designed a Clever Experiment

University of Tucumán, Argentina (2002-2003)

Sample

  • 330 university students
  • 33 sessions of 10 participants each
  • 5 workers + 5 employers per session

Task

Solving computer mazes (15-minute work period)

  • Practice round → projected performance
  • Actual production → revealed ability

Five Treatments Vary the Interaction Mode

Each Employer Rates 5 Different Workers

Treatment Resume Photo Phone Face-to-Face
B (Baseline) ✓ — — —
V (Visual) ✓ ✓ — —
O (Oral) ✓ — ✓ —
VO (Both) ✓ ✓ ✓ —
FTF (Face-to-Face) ✓ ✓ ✓ ✓

Each employer rates 5 different workers → Employer fixed effects

Randomization Ensures Exogeneity of Treatment

What the Experiment Identifies

What Randomization Gives Us

  • Treatment is exogenous by design
  • \(\text{Cov}(\text{treatment}, \varepsilon) = 0\)
  • No selection into treatment groups
  • Causal interpretation of treatment effects

What It Does NOT Give Us

  • Beauty is not randomly assigned
  • Cannot claim causal effect of beauty on productivity
  • But: Can test if beauty predicts productivity
  • If no correlation → productivity channel ruled out

Randomization of treatment ≠ randomization of beauty

Incentives Ensure Truthful Revelation

Both Workers and Employers Are Paid for Accuracy

Worker Payment

\[ \begin{aligned} \text{Pay}_j &= 100 \times \text{Actual}_j \\ &- 40|\text{Estimate}_j - \text{Actual}_j| \\ &+ \sum_{i}\text{Wage}_{ij} \end{aligned} \]

  • Incentive for high performance
  • Incentive for accurate self-assessment

Employer Payment

\[\text{Pay}_i = 4000 - \sum_{j=1}^{5} 40|\text{Wage}_{ij} - \text{Actual}_j| \]

  • Incentive for accurate worker assessment
  • Penalised for over/under-estimating ability

Productivity Does Not Explain the Premium

Beauty Has No Effect on Maze-Solving Ability

Productivity Channel Ruled Out (Treatment B Only)

\[ \widehat{\log(\text{actual})}_j = -\underset{(0.039)}{0.018} \cdot \text{beauty}_j - \underset{(0.012)}{0.039} \cdot \text{age}_j + \underset{(0.087)}{0.372} \cdot \mathbb{1}[j \text{ is male}] + X_{j}' \hat{\Gamma} \]

  • Beauty coefficient is not statistically significant (\(p > 0.10\))
  • Males solve 37% more mazes (\(p < 0.001\))

Beauty does not affect productivity

Confidence Transmits Through Speech

Beautiful Workers Report Higher Confidence

+17% Confidence per SD of Beauty

Residualized confidence (after controlling for actual and projected ability) vs. beauty. The positive slope indicates overconfidence.

+17% confidence per SD beauty, even after controlling for actual ability

Beauty Causes Overconfidence

Controlling for Both Actual AND Projected Ability

Beauty STILL predicts higher confidence. This is OVERCONFIDENCE

Confidence Only Affects Wages When Employers Can Hear

Oral Transmission Channel

Effect of log confidence on log wage by treatment. Error bars show 95% CI. Confidence only matters in treatments with audio interaction.

Confidence is communicated through speech, not appearance

Multiple Channels Contribute

The Beauty Premium Emerges with Interaction

From 0% to 12% Across Treatments

Effect of one SD increase in beauty on log wage by treatment. Error bars show 95% CI. No premium without interaction; premium emerges with visual OR oral contact.

No premium without interaction

Channels Are Not Fully Additive

The Full Model with Interactions

Decomposition coefficients from pooled regression. Visual and oral channels both contribute, but with negative interaction (channels overlap).

Face-to-face ≠ Visual + Oral (diminishing returns)

Visual ~40%, Oral ~40%, Confidence ~20%

Channels Have Diminishing Returns

\[ \underbrace{6.8\%}_{\text{Visual}} + \underbrace{11.3\%}_{\text{Oral}} + \underbrace{2.7\%}_{\text{Confidence}} - \underbrace{7.1\%}_{\text{Overlap}} \approx 13.7\% \]

The negative overlap term means channels have diminishing returns—face-to-face ≠ visual + oral

Employers Do Not Sacrifice Accuracy for Beauty

No Evidence of Taste-Based Discrimination

\[\begin{aligned} \widehat{\log(\text{wage})}_{ij} &= \underset{(0.031)}{-0.010} \cdot \text{beauty}_j - \underset{(0.055)}{0.010} \cdot \mathbb{1}[i \text{ sets } j\text{'s wage}] \\ &- \underset{(0.057)}{0.058} \cdot (\text{beauty}_j \times \mathbb{1}[i \text{ sets } j\text{'s wage}]) + X_{ij}'\hat{\Gamma} \end{aligned}\]
  • Interaction coefficient: \(-0.058\) with \(p > 0.10\) (not significant)
  • Employers do not inflate wages for beautiful workers when it “counts”

The premium reflects beliefs about productivity, not tastes for beauty

Laboratory Results May Not Generalize Directly

External Validity and Policy Implications

Limitations

  • Student subjects (may not generalize to professional workers)
  • Simple task (maze-solving vs. complex jobs)
  • Short-term interaction (one session vs. repeated interactions)
  • Laboratory setting (lower stakes than real hiring)

Policy Implications

  1. Structured hiring with standardized questions
  2. Phone screens before in-person interviews (removes visual bias)
  3. Interviewer training on overconfidence detection

Summary

From Correlation to Mechanism

Findings

  • Observational evidence: 15% beauty premium exists
  • Experiment: Mobius & Rosenblat (2006) isolates channels
  • Productivity: Ruled out (beauty ≠ ability)

Mechanism

  • Overconfidence: Beauty causes excess confidence
  • Oral transmission: Confidence only affects wages through speech
  • Decomposition: Visual ~40%, Oral ~40%, Confidence ~20%

Next Week: Exploiting time variation—Panel data methods

References

References

Hamermesh, D. S., & Biddle, J. E. (1994). Beauty and the Labor Market. American Economic Review, 84(5), 1174–1194.
Mobius, M. M., & Rosenblat, T. S. (2006). Why Beauty Matters. American Economic Review, 96(1), 222–235. https://doi.org/10.1257/000282806776157515